Work

Analisis Kerentanan Model CNN terhadap Serangan FGSM pada Klasifikasi Citra X-ray COVID-19

Year: 2026
Type: dissertation
Abstract: The advancement of deep learning, particularly Convolutional Neural Network (CNN) architectures, has brought significant changes in the automated diagnosis of lung diseases through chest X-ray (CXR) images, including pneumonia and COVID-19. However, despite achieving high diagnostic precision, CNN models remain highly susceptible to adversarial attacks, defined as small, visually imperceptible alterations that cause high-confidence mispredictions. This study aims to analyze the vulnerability of the DenseNet-121 model to Fast Gradient Sign Method (FGSM) attacks in classifying COVID-19 chest X-ray images. The research method involves training the DenseNet-121 model using the COVID-19 Radiography Dataset consisting of 21,165 images with four diagnostic categories: COVID-19, Lung Opacity, Normal, and Viral Pneumonia. The model was then tested on clean data and FGSM-perturbed data with varying epsilon values. The results show that at epsilon 0.1/255, the global model accuracy dropped drastically from 95.42% to 59.28%, with an Attack Success Rate (ASR) of 40.72%. Class-wise analysis shows that the Lung Opacity class experienced the highest F1-Score degradation of 51.13%, followed by COVID-19 at 48.83%. Gradient-weighted Class Activation Mapping (Grad-CAM) visualization confirms that this performance degradation is associated with a shift in the model's attention from clinically relevant pulmonary regions toward background noise (semantic decoupling). This study concludes that high accuracy on clean data does not guarantee model robustness against adversarial attacks, necessitating the integration of defense mechanisms such as adversarial training to improve model safety in clinical deployment.
Language: English
Open Access status: green
OpenAlex ID: W7219437683
Cites: 48
Cited by:
Related to: 10